priorlabs/tabpfn

TabPFN is a pretrained AI model designed to work on tabular data, meaning data organized in rows and columns like a spreadsheet. Most machine learning models need to be trained from scratch on each new dataset, which takes time and requires a meaningful amount of data. TabPFN takes a different approach: it is a foundation model that has already learned general patterns across many datasets, so it can make predictions on a new dataset in seconds with very little setup. The core use cases are classification (predicting a category, like whether a customer will churn) and regression (predicting a number, like next month's sales). You install the Python package, load your data, call .fit() and .predict(), and you are done. The interface is intentionally familiar to anyone who has used scikit-learn. The model downloads a checkpoint file automatically on first use. TabPFN performs particularly well on small datasets where there are not enough rows to train a large custom model. The research behind it was published in Nature. On larger datasets or when using it in production systems, a GPU is recommended. Without one, the tool only handles datasets up to about a thousand rows at a reasonable speed. A hosted cloud version called TabPFN Client is available for those without suitable hardware. The project includes a wider ecosystem. Extensions add features like SHAP-based interpretability, outlier detection, synthetic data generation, and support for problems with many output classes. A no-code web interface lets non-technical users try the model without writing any Python. The model weights carry a non-commercial license. The code itself is under Apache 2.0 with an attribution requirement. A commercial enterprise edition exists for high-throughput production use, with a distillation option that converts the model into a faster, lighter form.